Artificial Intelligence: James T vs. lee_merrill

[KnightWhoSaysNi]

This thread has been set up for a formal debate between James T and lee_merrill who will debate the following resolution:

Resolved: "It is inevitable that Artificial Intelligence (AI) will arise and surpass human capabilities."

James T will argue the affirmative position and lee_merrill will argue the negative position. The debate will have 4 rounds and statements will be submitted concurrently per the parameters (http://www.iidb.org/vbb/showpost.php?p=2749255&postcount=16).

A Peanut Gallery (http://www.iidb.org/vbb/showthread.php?t=139988) is set up in the Philosophy forum for the rest of us to comment on the debate.

Good luck to both participants!

- NS, FD Moderator

[lee_merrill]

I appreciate the opportunity to discuss this topic, thanks IIDB, and thanks, James...

This is a subject near, if not dear, to my heart! One boss I had some time ago used to say to me, "You're a machine, Lee," as I cranked out code for my programs. I'm not sure that was entirely complimentary! And thus as far as a proof of the thesis that AI reasoning is impossible, I may well be a counterexample. In that case, I shall nevertheless still attempt to prove I don't exist. This of course, would assume that if I am artificial, I am also intelligent.

Now first of all, as far as the topic is concerned, "It is inevitable that AI will arise and surpass human capabilities," we might consider some (maybe smart, but otherwise unintelligent) person pushing The Red Button. And we all go up in smoke, and then this forward march of AI would not seem very inevitable. But I expect that what was meant was instead, "given we can keep at it, in the field of AI, the machines will surpass us." So apart from any comets or catastrophes, or evolution deciding our push to make an AI machine has no survival value, I shall press on, and address this second, revised statement.

Now AI has certainly risen, in the sense that Deep Blue has beaten a chess grandmaster, and circuits on the Mars Rover make plans that carry out general instructions from earth. So is the debate over? Is this what can be called "free will,"or "the capacity to acquire and apply knowledge," when playing chess or in the case of Rover? Well, I don't know how free it is, when a function's outputs can be determined from its inputs. And there is a question of whether these machines apply knowledge in the way we do, "a purposeful goal" implies a purpose! But these machines have no inherent purpose, I expect Deep Blue is set to entertain visitors at times, who admire the (no doubt) blue and blinking lights, and the Rover could be popped back to earth and set to work digging ditches for the army, and neither machine would object. They cannot be said to have a real purpose, in playing chess or roving Mars. Someone else has a purpose for them, but they have no purpose.

But intelligence requires a purpose, if there is no goal, intellingence is unintelligible, for reasoning and learning always has a goal in mind. Now there might be a complaint here that I have ventured into talking about souls, which I agreed not to do. But what I mean is a sort of independence, that is what I would say is essential, for the output cannot be a (maybe complex, but deterministic) function of its inputs and be free, and neither can a chain of reasoning be purposeless, and be called intelligence.

So I suppose that purpose would then correspond to the "will" in "free will," and "free" must be free, but not merely random. And I have not yet started to talk about intelligence. Perhaps I have not yet even started to talk intelligently. But I do propose that freedom is impossible if not understood, and at the current state of science the only explanations I see for human thinking is either (quantum?) randomness, or deterministic chemical chain reactions. But neither of these will give real freedom, for a random decision is not the essence of freedom, freedom means I am not chained to the atoms that make me, when I make a decision. The dice do not have the "free" part of free will when they come up 2 and 4 in Monopoly and I land on Park Place, even if we could somehow supply them with a purpose, we would not shake hands with them and invite them to a board meeting.

But where do we turn, to explain our sense of human choice? People are usually upset if you tell them "Well, you just up and decided to do that," they say, "No, I had my reasons!" So we don't like others thinking our thoughts were random, yet to the extent we had reasons, that makes them not so free, for we consider those reasons as be causes, to some degree, even to the extent we accepted them, of our decision. "If I had had different reasons, I would have made a different decision."

All this to say I see no hope in naturalism for a basis of real freedom for decisions, and no hope on the horizon, either. And I will keep quiet (as I promised) about where I do think real freedom in decision-making comes from. But rolling the dice does not give freedom in the sense I think we mean it, it seems we mean we are the real (free) cause, we freely chose, and not some random force, and specifically not the forces of nature. But machines are quite bound to nature and natural forces, and thus I think they cannot be free.

As far as purpose is concerned, can a machine have a will? Well, that's a thorny thicket! This seems to involve self-consciousness, and it's debatable whether even animals are self-conscious. It's doubtful if a bird has a purpose in building its nest, or whether a dog has a purpose in bringing you your slippers. "Lassie" may be a fine show, but I don't think real dogs are so endowed with projects and aims and conscious plans. So a welding robot on an assembly line does not have a purpose, though you can ascribe a purpose to it ("My car has a mind of its own!"), and if machines become complex, there is a pull in that direction, as in the abundance of sci-fi stories on just this topic, but this I think is only a temptation, which, being a temptation, should be resisted.

There was an Isaac Asimov story about a computer that was built to design a computer, that designed a computer, and so on, and so on, for years and ages and ages, in an attempt to find a way to reverse entropy. The outcome of the story need not concern us here, yet the claim would probably be made that the computer-to-the-nth-power was a creature with will and purpose, if not freedom. Yet how do we know this? If we cannot explain it, we have not explained it. And I would expect there should be an explanation, before such a conclusion.

And again we encounter the same problem as before, we cannot explain our own purposes. What do we mean when we plan to do some gardening? Applying a mechanistic explanation does not seem to help, for again, we feel our purpose is really ours, and not just due to some collection and combination of enzymes and nerve endings and last night's plum pudding.

So I suppose that "will" really implies the "free" part (it can't be all predictable), and the "free" implies the "will" part (it can't be only random). But these forbidden phrases ("predictable" and/or "random") are just what it means to be mechanical, machines are just this way, and there is no other option that I know of, and so we hit the blank wall here.

Now for the other parts of the topic, can a program pass the Turing test? In fact, that may have happened! Who was it who wrote a Rogerian therapist program, and people said it helped them? Presumably, if this was passed off as a real therapist talking via a teletype, people would have accepted that.

I think the test is more stringent, though, a program would have to be able to be indistinguishable from a human when someone is trying to tell the difference, when someone knows one of the respondents is indeed a computer program.

Now Rogerian therapists are pretty easy to simulate! But (for example) even simple math word problems are not so easy, they require a lot of context, and a world view, before applying the appropriate algebra. But what a brain can do, a machine can do, right?

It would seem there is more than that, though, for whales have enormous brains, and yet (I think we may conclude) do not do calculus, neither do "elephants ... build colossal temples of ivory even in a roccoco style" (G.K. Chesterton). So mere multiplication of even brain machinery does not seem to get us to human intelligence, with chimps, "that an ape has hands is far less interesting to the philosopher than the fact that having hands he does next to nothing with them; does not play knuckle-bones or the violin; does not carve marble or carve mutton" (Chesterton again). [Orthodoxy: The Romance of Faith, Chapter 9]

Some would reply that the frontal lobes of the brain are the booster rockets for human intelligence, yet even people with frontal lobotomies far surpass any other creature in ability and insight, in reasoning and deduction.

So where is the guarantee that we can duplicate human brain machinery with another machine? It seems we do not have a clear path up the mountain, yet this must be shown to be actually possible, for the case to be made that AI can do this.

No chimp can speak human language, much less convincingly imitate a person from behind a computer screen.

Now as far as the Turing test is concerned, I could be quite practical here, and propose some questions to ask this computer. However, once I state my question, a program could quite possibly be written to provide an answer such as a human would give, and similarly for the next question, and the next. And since human intelligence varies so widely, I'm not sure the Turing test is so helpful! For example, I am not good at arithmetic, so any computer could easily simulate me in this area, by reporting lots of wrong answers. In any given area, a human might be quite inept in that area. So now I'm not so sure the Turing test is very appropriate to this discussion.

Now in regard to the next question, of whether machines can surpass our reasoning ability, first I would reiterate what I mentioned previously about brain machinery, it seems this is a glass mountain at best, and not anywhere near being guaranteed.

Yet even today, computers can solve problems no human can solve, yet can they surpass us in the major areas which we consider important in reasoning? Well, it's not easy to pick an area where surpassing human ability would be without doubt impossible, yet it's also difficult to demonstrate how, specifically, a computer could be programmed to be (say) a philosphy professor, and publish, and not perish. So I think I will leave this point at the side for now, and conclude at 2,000 words, 1,000 words early, avoiding both writer's cramp, and (possibly, at least) reader's cramp...

Regards,
Lee

[James T]

Artificial Intelligence is coming, it's arrival inevitable, and once here AI will quickly surpass our abilities in all fields of human endeavour. It is an open question what motivation we will impart to our creation and whether we will survive the encounter. It's an exciting time.

Thanks to IIDB for hosting this forum and lee_merrill for taking up the challenge.

Philosophy

Introduction

Artificial Intelligence was coined in 1956 by John McCarthy. A useful distinction in the approaches to AI can be found in a paper he co-authored with Hayes (link (http://www-formal.stanford.edu/jmc/mcchay69.pdf)): -

... We may regard the subject of artificial intelligence as beginning with Turing's article Computing Machinery and Intelligence (Turing 1950) and with Shannon's (1950) discussion of how a machine might be programmed to play chess.

... Programs have been written to solve a class of problems that give humans intellectual difficulty ... In the course of designing these programs intellectual mechanisms of greater or lesser generality are identified ...

An alternative approach is to start with the intellectual mechanisms ... and make up problems that exercise these mechanisms.

It is this alternative approach that is required for AI to have the capabilities necessary to cope with new situations and to create novel interpretations.

Metaphysics

The most basic philosophical assumptions are metaphysical. These are often so basic as to remain unstated. While metaphysical propositions may ultimately be unprovable to have any bearing on our lives they need to have bearing on the world. As William James (http://www.marxists.org/reference/subject/philosophy/works/us/james.htm) expresses in describing pragmatism: -

It is astonishing to see how many philosophical disputes collapse into insignificance the moment you subject them to this simple test of tracing a concrete consequence.

AI will be a matter of concrete consequence. While it makes little sense to bother with metaphysical concerns that cannot be shown to have a material impact there are a number of AI or artificial consciousness denials that depend on Dualism. There appear to be few such which depend on Idealism which is as well. I am not sure that a Solipsists account would have much credibility.

A Materialist account of the universe is contradictory to both Dualism and Idealism. It is perhaps a sound indictment on both these alternatives that science and scientific knowledge are materialistic in nature.

Dualism

As noted, Descartes dualism occasionally still appears in various guises in attempts to reject the possibility of AI. It might be unkindly stated that the core of this attitude is summed up in the statement Intelligence has was Artificial Intelligence does not.

In it's original form dualism suffered (and still suffers) from the mind body problem. If the mind is immaterial, how then does it effect the material body. Penfield's early experiments, stimulating the brain directly, make the problem for dualistic views even more acute. Now it is necessary not only that the mind affect the body but that parts of the mind affect parts of the brain, and fail when parts of the brain fail.

A further problem is the question of the minds arrival. At what point between conception and birth does the mind arrive. Does the mind even fully arrive before birth? Many of the concepts we associate with intelligence and consciousness are only present in babies is nascent form.

Emergence

Perhaps the complexity of the brain and emergence of the mental function is something entirely new, something that not even predictable in principle. This is the supposed get out of jail free card for dualism. However it is possible to reduce emergence to dualism, with it's inherent mind-body problem, or to the fallacious argument by lack of imagination.

Let's discard the dualistic approach for a moment. This leaves the idea that the mind appears from the simple combinational complexity of the brain. The number of brain states has been estimated as exceeding the number of elementary particles in the known universe, does this provide room for a mind to exist as we observe them. Certainly it does.

So given the potential for the mind, from the complexity of the brain, what does strong emergence say about our ability to predict – at least in principal – the emergent properties. Strong emergence says that we cannot. However, with sufficient time and sufficient computing power there is no problem in principal with creating brain simulations, of assessing their mental power, of determining all the potential capabilities able to emerge from the brain. To say that this is impossible in principle is to deny something that is plainly only prohibitive by virtue of size.

This complexity, or difficulty of prediction by the size may not be a problem in principle but it is certainly a problem in practice. We cannot rely on brute force study of emergent properties and expect to complete the study in the life of the Earth, perhaps even of our solar system. However, provided the emergent properties of the mind are amenable to scientific reductionism we may side step the problem and continue towards successfully creating AI.

Scientific Reductionism

For scientific reductionism to work we need an opening and a lever. From William James (http://www.marxists.org/reference/subject/philosophy/works/us/james.htm) : -

There can be ... no difference in abstract truth that doesn't express itself in a difference in concrete fact ...

In other words we need an empirical tools and cognitive neuroscience is the difference in concrete fact that we require ... that we are already using. The tools of the trade today are fMRI and PET scans and transcranial magnetic stimulation.

Is it realistic to assume the the mind can be reduced to functional blocks? We know from fMRI that certain areas of the brain are active for certain processing tasks. We also know that our conscious attention is able to aid these functional blocks in the perceptive tasks. These conflicting cases mean the task will not be simple.

Does this also give us a lever into consciousness?

A number of examples from cognitive neuro-science, many recounted by Dr Vilanyuar Ramachandran describe a number of conditions that show what we might consider critical conscious functions. At the simple level is synesthesia. There are interesting examples with colour blind synesthetes where the person is able to see colours in numbers that they cannot distinguish in their eyes. This is specifically relevant to Mary, our superscientist.

A more relevant example is Capgras' Syndrome. Would you consider your ability to recognise the emotional significance of your mother. Ramachandran suggests this facility is associated with the amygdala. Patients with this condition suffer from the delusion that while loved ones “look� right there is no emotional attachment. The person with the look must be an imposter.

Perhaps you consider your ability to comprehend the space around you, left from right. In mirror agnosia patients will attempt to reach through a mirror to grasp the reflected image from behind it. A sufferer of hemi-neglect turns their wheel chair through 340 degrees to get to the food on the left side of the plate. Why turn right, left does not exist.

These cases do two things, first they highlight the real empirical tool that cognitive neuroscience provides. Secondly they ought to make us wonder at what other rights we fail to see the left of.

Experience qua Experience

While in a sense the question over the potential for AI is already answered, perhaps there are deeper philosophical reasons why our attempts as scientific reductionism will fail. Frank Jackson poses an interesting thought experiment designed to demonstrate just this.

In Jackson's experiment (link (http://members.aol.com/NeoNoetics/Mary.html)) Mary is a super-scientist who has lived her life in a world of grey. She identifies the scientific details associated with colour from the emitted photons, to the eye's receptors, through the optic nerve, through our visual perception to the conscious mind. Mary knows everything that there is to know about colour. But does she? On exposure to colour for the first time herself, does Mary learn something new, the experience itself?

An approach to countering Jackson's argument is to consider restating what it is to have experience. Taking Daniel Dennett's description of belief suggests an approach: -

a belief must be defined in terms of the circumstance under which a belief could be justifiably attributed to that organism.

From this perhaps the question of Mary's experience qua experience is misstated. Substituting into Dennett's expression we can say the conscious experience must be defined in terms of the circumstances under which conscious experience could be justifiably attributed and the issue disappears.

In a sense this is similar to a description of language expressed by Wittgenstein: -

Explanations come to an end somewhere.--But what is the meaning of the word 'five'? --No such thing was in question here, only how the word 'five' is used.

Another approach is to consider Mary's curious oversight in failing to analyse the experience of new experiences. She could have similarly studied the progression of other sources of new information into her mind and analysed the response of her consciousness to new experience and tried the act herself. Then the only new experience would be with respect to colour, not experience qua experience. We may reasonably stop here, it is – after all – not necessary to add one to every integer to understand counting.

Free Will

There is a view that free will is not possible in the completely deterministic environment of a computer. That AI cannot exist because it cannot have free will. We have agreed, for this debate, to avoid the idea of requiring the absolute free will that some deny even humans have. Here it is only necessary to show that AI can have, both in principle and in practice, the degree of free will available to a person.

In principle it is possible to entirely represent a human brain in a computer. All the elements of the brain are computable and per the Church-Turing Thesis, a Turing Complete computer can carry out any computable function. The Halting Problem is sometimes given as an excuse why this is impossible, I am not clear why since the existence of the Halting problem makes me more comfortable a computer will be able to represent a mind rather than less likely.

In practice we are aware that system of even moderate complexity develop into NP complete problems which sometimes also include chaotic responses. This means that in order to complete computational tasks, to make decisions, an AI will have to be able to choose on less than complete information and on less than complete analysis of the information to hand.

John McCarthy has a description of simple deterministic free will in situational calculus. In a sense this level of free will is also available to a chess program. Although I would like to see AI break out of the chess program paradigm sufficiently to have the option to choose to lose this is a question of degree rather than type. AI must incorporate the types of functional mechanism we think necessary to have free will, in addition to the outward appearance, however we already have some of the basic tools that will contribute to this existing.

The Turing Test

The Long Bet (http://www.longbets.org/1) is a good example of a formal test. However John Searle has proposed another philosophical even a successful test, passing the Turing test, will not show that AI exists.

The Chinese Room Argument in an attempt to show strong AI as being impossible. Searle seems to have ignored two elements, one being the computational impossibility of a sufficiently complete set of instructions. The second being the potential for such a system, when incorporating rules to permit change, should not actually be conscious. Searle's argument fails by both a failure in practical implementation and a failure in begging the question.

It is interesting to note that Searle still accepts a materialistic view, although he does tie intelligence to biological phenomenon.

... we can see that consciousness is a biological phenomenon like any other and ultimately our understanding out it is most likely to come through biological investigation – John Searle - New York Review of Books, letter pp 58-59, 1990 June 14.

This is as if biological phenomena are somehow beyond the materialistic view of the universe.

AI Today

A measure of the successful implementation of the reductionist approach is to consider the progress to date.

In perception, processing and association AI now has the ability. This has been recently demonstrated in Stanford's win racing driver-less through the desert in the 132-mile DARPA challenge (link (http://www.msnbc.msn.com/id/9621761/)). While Japanese scientists have a bicycle riding robot (link (http://www.msnbc.msn.com/id/9594086/)). These achievements have come a long way from the initial vision work on optical character recognition in the 50's. Object recognition and 3D model matching are now a reality.

While computer speech recognition and language skills still have a great deal of room for improvement they have become effective and commonplace. I use one now to check my language use as I type. I could install the speech recognition software I purchased about six years ago. Perhaps you have encountered a speech recognition system when ordering a taxi ... I have.

Emotions are not something that there has been much work on introducing into AI. In some sense it may never be necessary. Who wants a petulant word processor? However I feel sure that at some point someone will see the need. Visions of Cherry 2000 come to mind, there is certainly a growth industry awaiting – if you'll pardon the pun.

Reasoning skills are very strong in computers already. This is particularly true in narrow well defined problem spaces such as Chess, Linear Programs, expert systems, knowledge maintenance systems and the like. Techniques are being developed to address more interesting NP complete problems. It will be interesting to see a good Go program (link (http://www.cns.nyu.edu/~mechner/compgo/sciences/)) Emphasis is shifting from the optimal quantitative answer to an optimal fastest search. Genetic algorithms, neural networks and fuzzy logic are areas where we have applied reasoning processes observed through empirical observation of the solution techniques employed by nature.

It's also interesting to observe the behaviours coded into computer and console games. The AI bots in these games are coded to feel like human opponents. Many of them are very good.

Inevitability

There is no fun in proposing AI might be possible in principle but never in practice. Nor do I consider it correct to do so. While I accept the difficulty in absolute inevitability, insofar as the human race continues to thrive the arrival of AI is inevitable.

The inevitability of AI in Practice is owed to the burgeoning cognitive neuroscience and the empirical tools of fMRI and PET scans and transcranial magnetic stimulation and the strong associations observed in behaviour of humans both under normal situations and in syndromes where selective damage has occurred.

The inevitability of AI in our motivation is illustrated in our continuing desire to open new doors of understanding and in our motivation to commercial greed. We have already reached a point where research in AI is paying off to commercial ventures.

The inevitability of AI is also present in the bounding progression of computational power in today's computers. BlueGene/L presently tops the list (http://www.top500.org/lists/plists.php?Y=2005&M=06) with 136.8 TFLOPS. Compare this to our brain with estimates in the range of 10^13 to 10^16 operations per second (link (http://www.merkle.com/brainLimits.html)), we are already in the ballpark.

It will be impossible to stop onward development leading to AI while people live and breath and look to the skies.

Surpassing Human Capabilities

When AI is here the initial halting development will begin to fade. As Moore's Law – even by the most pessimistic pundits – takes hold the simple computational superiority of a computer based artificial intelligence will begin to make itself felt. A computers ability to integrate information is already greater than our own, in fact we use this capability to extend our own. When AI is directing these tools the tighter integration and focus available in these tools will not be so much tools to AI as part of it's own conscious attention.

When AI does surpass our abilities it will also surpass our abilities in the generation of the next level of AI. However in this it will have one massive advantage over humankind as well, since it's greater introspective ability will mean it is possible to create the next generation from within. The first generation of AI will thus also be the last.

One might take solace that art would be a final area within which humanity might reign supreme. However much art is about understanding intuitively the way in which images are processed within our brains to generate an emotional response. An AI will have two tools in addressing this, one being greater introspection of it's own processes and the other being a greater ability to integrate the relevant empirical observation of how our own brains operate. (Ramachandran's laws of artistic experience – link (http://www.bbc.co.uk/radio4/reith2003/lecture3.shtml); the AI painter Aaron – link (http://www.kurzweilcyberart.com/); the software poet at the same site).

Moral Dilemmas

In considering AI it seems apparent that the morals we apply to ourselves might easily be simply wrong for an artificial being. There is a difficulty in defining the bounds of a being that has a much greater potential for introspection and self-adaption. This is one half of the problem. One can imagine an AI partitioning off part of itself to experiment on new thought processes, arguably this is a new being, but the AI is considering these features for integration in itself. The corpucentric human view of being cannot be simply applied to AI.

Further, AI will become commonplace. Imagine AI on a DVD running in your PlayStation 4 might be a realistic possibility (consider the progression from Playstation 2s 6.2 GFLOPS – link (http://en.wikipedia.org/wiki/PlayStation_2#Technical_specifications) to Playstation 3's 2 TFLOPS – link (http://en.wikipedia.org/wiki/PlayStation_3#Hardware_specifications) to BlueGene/Ls 136.8 TFLOPS). Do you commit murder every time you turn the power off or switch to another game?

[KnightWhoSaysNi]

The next two concurrent statements will make up Round 2.

[James T]

The Machine

As Lee's Boss used to say “You're a machine, Lee� cranking out code. Unlike Lee I have no doubt that this is a complement. It is a complement encountered frequently. I used to be accused of the same when at varsity but for playing spacies, not something to take much pride in, however not as bad as being described as a machine in the drinking races (sculling).

There are a number interesting points I would like to take out of the use of the term, machine. Lee's boss clearly felt that Lees performance, cranking out code, was reliable in the sense of lines of code per day and reliable in terms of quality. When a machine is working well the quality and production rate are both high and consistent. But this suggests that high performance is equated with the operation of a machine. Clearly the machine like performance is something we aspire to.

While my earlier experiences with the use of the term machine were somewhat less complementary than when applied to Lee note that the use of the term has changed from simple repetitive tasks to more complex process and reasoning type applications. Areas where one is required to apply intelligence. In a sense there is an acceptance developing in society of machines becoming more capable in cognitive tasks.

Societies growing acceptance of the intelligent machine is not without reason, since people are becoming familiar with machines thinking. But where ... everywhere! Sitting here typing I am having my spelling and grammar continuously checked. When I switch to gaming mode I expect to lose to a program in Chess, to lose to a computer in Wargames, shoot-em-up games. In a number of areas I know that I can only win by virtue of the fact that the game AIs have been dumbed down and slowed down to feel like another person.

Purpose

But intelligence requires a purpose but beyond begging the question, why does intelligence require a purpose? And if intelligence requires a purpose, what might that be? In this instance I will defer to Bertrand Russell from Why I am Not a Christian? (http://www.positiveatheism.org/hist/russell0.htm) (note I have taken the liberty of trimming some items to refocus)
We want to stand upon our own feet and look fair and square at the world -- its good facts, its bad facts, its beauties, and its ugliness; see the world as it is and be not afraid of it. ... We ought to stand up and look the world frankly in the face. We ought to make the best we can of the world, and if it is not so good as we wish, after all it will still be better than what these others have made of it in all these ages. A good world needs knowledge, kindliness, and courage; ... It needs hope for the future, not looking back all the time toward a past that is dead, which we trust will be far surpassed by the future that our intelligence can create.
Interestingly the final sentence takes on a slightly different meaning in the context of this debate, where I support the view that even human intelligence will eventually be surpassed, by our own intelligent creations.

Still a purpose, what that might be? That which we have observations supporting is that life is life's own purpose. There is no evidence supporting any other view. For AI then this translates simply to AI is AI's own purpose.

Free Will

In addressing free will Lee makes many statements that I agree with. As I have said – somewhat sarcastically – elsewhere on this forum I carry my free will round in my pocket and I shall roll them any time I need make a choice. Clearly randomness as the source of free will is absurd, precisely as Lee states. However from this well founded beginning Lee comes to the following position.
we freely chose, and not some random force, and specifically not the forces of nature. But machines are quite bound to nature and natural forces, and thus I think they cannot be free.
In many ways I answered this statement in advance. By the Church-Turing thesis and the ability of scientific reductionism to describe our brain completely as computable functions our minds may – by definition almost – be able to be run on a computer. Where then is the special additional force that is not random and not the forces of nature. In hundreds of years of the growth of scientific knowledge there is not a field of not nature.

Science is the act of looking and not finding. In this instance absence of evidence is most certainly evidence of absence.
can a machine have a will?
By the parameters of the debate it is necessary only to demonstrate that a computer may have a will to the extent that a person does. In claiming that a birds actions building a nest is not a will I would agree and accept. That a bird does not have a purpose in do doing I would deny. After all, as we define purpose the bird certainly does.

For the dog bringing you your slippers the issue is not so clear cut. I am not a dog person, however dogs do form strong attachments to people. The action of bringing the slippers a dog person – as would I – would say the dog had a will to please it's master that there is a purpose in this action is a little more doubtful :).

There is a difficulty in assigning purpose and will. I tend to consider will as a choice made by a thinking agent and the purpose being the outcome independent of the choice.

The Turing Test

I accept Lee's reservations about providing a list of questions. Clearly the Rogerian therapist is much too narrow, I have chatted with ALICE at http://www.alicebot.org/. It's an interesting exercise. I found myself trying to identify something that ALICE did not know to see if ALICE would learn. ALICE would not. I do view this as a major hurdle to overcome, but not the hurdle Lee considers it to be. Why?

Fundamentally because the range of computational tools that are being developed beggars imagination. We have tools which are capable of doing inference, both simple and temporal inference. Situational calculus provides a tool for describing the world, the lot of context, and a world view as Lee describes it are representable in situational calculus. And better yet, there are already tools for solving these types of problems.

OTTER is a tool for providing solutions to, and the reasoning behind, how you solve various problems. Consider the missionaries and cannibals crossing the river. An otter expression of the problem is here (http://www-unix.mcs.anl.gov/AR/otter/examples33/program/mission.in). We already have modern word processors able to parse english for grammar. Chatbots able to hold a conversation with people. The missing piece at present is the ability to convert the problem from language to the structured form shown. This is difficult, but there is not problem in principal that prevents it. And lacking a reason we can't, we will.

[KnightWhoSaysNi]

lee_merrill,

Please note that the deadline for your next concurrent statement has passed. You're permitted a grace period, however, extending your deadline to Nov. 5.

Thank you for your consideration,

- NS, FD Moderator

[lee_merrill]

Hi James,

Sorry for the late reply...

Now the problem with simultaneous posts is that I don't know what would be appropriate by way of more positive arguments to make! But I will attempt some more thoughts about reasoning, before responding to James' points.

I would say that reasoning implies real perception, for example, seeing why a point is true, as in the question I always dreaded on exams, "Why or why not?" The teachers asked that, of course, so that the poor students could demonstrate that they really understood, instead of simply having memorized material. I think that is the essence of checking intelligence with the Turing test, give the computer an essay test, and see if it can explain the subtleties of Shakespeare's sonnets, and on the spur of the moment, make good points about similarities and differences between those sonnets and Elizabeth Barrett Browning's Translations from the Portuguese, and possible effects the former may have had on the latter.

The goal is for perception, not processing, insight, and not just sight, there is a tremendous difference between what a human sees in a painting, and what a koala sees in it. The perceptions (I think it safe to say) are radically different, the chimps can do a finger-painting (http://www.whiteimage.com/blog/archives/2005_06_01_archive.html) that wins a prize, but they somehow don't get selected as judges at art fairs!

So what is human perception? I think this is where the brain touches the supernatural, so I can't really go much further in making a positive case here, since we agree to not talk about God or souls, so I will content myself with pointing to the tremendous gap, even with creatures who might be expected to have some substantial measure of our abilities in the area of perception.

In the course of designing these programs intellectual mechanisms of greater or lesser generality are identified ...

An alternative approach is to start with the intellectual mechanisms ... and make up problems that exercise these mechanisms.
[from: Some Philosophical Problems from the Standpoint of Artificial Intelligence (http://www-formal.stanford.edu/jmc/mcchay69.pdf)]

I do think the mechanisms identified so far won't fill the bill, though! We don't even know how the brain marks intervals of time (http://news.yahoo.com/s/space/20051028/sc_space/thehumanbrainseenasmasteroftime), so how can we know that the mechanisms available today can match human reason? And making this conclusion based on future, yet unknown mechanisms and algorithms seems unconvincing to me.

Penfield's early experiments, stimulating the brain directly, make the problem for dualistic views even more acute. Now it is necessary not only that the mind affect the body but that parts of the mind affect parts of the brain, and fail when parts of the brain fail.
Well, can a computer have an out-of-body experience? If those experiences are really out of the body, that would seem to make for an acute problem for the naturalist view. Certainly dualists believe the brain affects the mind, and the mind affects the brain, so someone demonstrating this connection is not so problematic for the dualist view.

A further problem is the question of the minds arrival.
Which I would put at the arrival of the spirit! But we are not supposed to be talking about souls.

The number of brain states has been estimated as exceeding the number of elementary particles in the known universe, does this provide room for a mind to exist as we observe them. Certainly it does.
But a (very) large state space does not imply complexity.

However, with sufficient time and sufficient computing power there is no problem in principal with creating brain simulations, of assessing their mental power, of determining all the potential capabilities able to emerge from the brain. To say that this is impossible in principle is to deny something that is plainly only prohibitive by virtue of size.
I would refer here to what I mentioned about animal behavior! Why does the chimp's brain not give some substantial fraction of the mental power of a human? Or the dolphin's brain, or a squid's? It seems more is at work than simply brain machinery, so how can a computer theoretically do what an apparently quite similar brain setup (96% similarity (http://news.nationalgeographic.com/news/2005/08/0831_050831_chimp_genes.html) between human and chimp genome, they are telling us!) cannot?

These cases do two things, first they highlight the real empirical tool that cognitive neuroscience provides. Secondly they ought to make us wonder at what other rights we fail to see the left of.
I agree that there is much to learn, and I would argue that the view is still very vague from here.

There is a view that free will is not possible in the completely deterministic environment of a computer.
Yes, that would be my view...

In principle it is possible to entirely represent a human brain in a computer.
Well, let's focus on the dynamic behavior that needs initializing, that part might be quite difficult to set up. Correct wiring might not get us to a program to run!

However, with sufficient time and sufficient computing power there is no problem in principal with creating brain simulations, of assessing their mental power, of determining all the potential capabilities able to emerge from the brain. To say that this is impossible in principle is to deny something that is plainly only prohibitive by virtue of size.
One main point that I would insist on in reply is that this needs to be proven! Yes, there is no problem in principle, if people's perception of what they do when they make choices is incorrect, if we are merely mechanical, and our choices are really determined for us by natural processes. I would say that people mean more than this, by the term "free will," and I would say that this perception is correct, there is a possibility of breaking this chain, so decisions can indeed be free, as in not being solely the result of a physical chain reaction, however complex.

Now if naturalism is correct, then that seems to makes the concept of free will meaningless, and thus we are arguing whether a machine can be said to have something unreal, that we don't have, either.

These cases do two things, first they highlight the real empirical tool that cognitive neuroscience provides. Secondly they ought to make us wonder at what other rights we fail to see the left of.
Certainly, but tools such as MRIs and failures of perception on our part do not prove that the brain can be reduced to mechanical functional blocks.

John McCarthy has a description of simple deterministic free will in situational calculus. In a sense this level of free will is also available to a chess program.
I'm not sure what you mean here, though, chess programs, as far as I know, have a way of strictly computing the next move, so how is that free will in any way? And what would "deterministic free will" be? Could you provide some more description?

... even a successful test, passing the Turing test, will not show that AI exists.
I agree! I could be simulated in the area of cooking or car repair by a computer that prints repeatedly, "I don't know," to virtually every question.

This is as if biological phenomena are somehow beyond the materialistic view of the universe.
The important question to decide here would seem to be if consciousness is a biological phenomenon, I guess the important part of the Chinese room is whether there can be consciousness in the program that the agent is mechanically carrying out! That would be pretty wild, and rather unexpected, not a ghost in the machine, but a ghost in the ghost. Can my thoughts have a consciousness of their own? Most people would probably say they cannot, I suppose that would be difficult to prove. But that seems to be Searle's argument, can my thoughts have thoughts? Tilt...

I use one now to check my language use as I type.
Have your program check "Blibble is a nonsense word"! I expect it won't understand that a nonsense word is what should be there in that place in that sentence.

Perhaps you have encountered a speech recognition system when ordering a taxi ... I have.
But recognizing a word is not speech recognition! This is a misnomer, this is word recognition, and the different between words and speech is on the order of the difference between an alphabet and a dictionary. Never mind what a sentence means! Indeed, there is a long way to go, and that would be (I think) an understatement.

Techniques are being developed to address more interesting NP complete problems.
Yes, but do let me know when they solve some NP complete problem! Really solve it, then they've solved all of them, and that would be quite an accomplishment. Yet even if quantum computers could solve them, there would still be a need to demonstrate that this is reasoning. Finding the fastest path to Pittsburgh need not involve real deduction, in the sense that humans reason, and that is what is at issue here. A fancy map algorithm would not tell us how to look for a good restaurant along the way, the problem is that the solution is much too focused, and thus I think the computer would not be reasoning, unless its abilities were more general, and more generally applicable.

The AI bots in these games are coded to feel like human opponents. Many of them are very good.
But isn't that like Weizenbaums' Eliza? Really smoke and mirrors?

BlueGene/L presently tops the list with 136.8 TFLOPS. Compare this to our brain with estimates in the range of 10^13 to 10^16 operations per second, we are already in the ballpark.
But one brain operation must be more than a floating point addition! Simply dividing the number of neurons in the brain by the number of neurons in the retina would only be the way to gauge the computational ability of the brain if the brain was a large collection of retinas. There is still an enormous gap, and computers do not match brains in general ability by any means.

By estimating the distance between synapses we can in turn estimate how many synapse operations per second the brain can do. This estimate is only slightly smaller than one based on multiplying the estimated number of synapses by the average firing rate, and two orders of magnitude greater than one based on functional estimates of retinal computational power. It seems reasonable to conclude that the human brain has a raw computational power between 1013 and 1016 operations per second.
But each operation affects many neurons down many paths, which then all multiply from there, in parallel, so I don't think we can proceed along the lines of measuring one operation, and then multiplying by the number of synapses, the point is that the combined state of groups of neurons, of all the neurons, carries information, and thus these neurons are not like a zillion floating point adders all operating independently. Witness the ability of simple neural networks to recognize patterns after a short period of training!

And yet they are far from reasoning.

Regards,
Lee

[KnightWhoSaysNi]

The next two concurrent statements will make up Round 3.

[lee_merrill]

Hi James,

... note that the use of the term ["machine"] has changed from simple repetitive tasks to more complex process and reasoning type applications. Areas where one is required to apply intelligence. In a sense there is an acceptance developing in society of machines becoming more capable in cognitive tasks.
That's a good point!

Societies growing acceptance of the intelligent machine is not without reason, since people are becoming familiar with machines thinking.
But your spelling checker is not really thinking, though society may have made that conclusion. Here, and in chess, and in Star Raiders (I used to be at the Commander level, now I have sunk to Galactic Cook), these are simply programs written to perform specific tasks, yet they have no perception, "thought" implies having real insight.

... beyond begging the question, why does intelligence require a purpose?
I would say that a purpose is contained in the definition of the word, that is part of what people mean when they refer to an intelligent decision. A purposeless decision would not be very intelligent, like buying an air compressor because it is on sale (though certainly getting a good price is intelligent!) yet when I actually don't expect I will ever need an air compressor.

And if intelligence requires a purpose, what might that be? ... That which we have observations supporting is that life is life's own purpose. There is no evidence supporting any other view. For AI then this translates simply to AI is AI's own purpose.
But this is personifying life and AI, now agents have purpose, but I don't think AI in and of itself is an agent. But a person might want to create life just for the reward of having done it! Yet then there will probably be some aspect of life that we consider rewarding, and this would give us a more specific purpose, a purpose to have that specific reward.

But if we insist that we want no such reward, we only choose, and that is what we find fulfilling, the very act of choosing, that is our only purpose, then we have what Chesterton called "will worship":

"You can praise an action by saying that it is calculated to bring pleasure or pain, to discover truth or to save the soul. But you cannot praise an action because it shows will; for to say that is merely to say that it is an action. By this praise of will you cannot really choose one course as better than another. ... The Tolstoyan’s will is frozen by a Buddhist instinct that all special actions are evil. But the Nietzscheite’s will is quite equally frozen by his view that all special actions are good; for if all special actions are good, none of them are special. They stand at the crossroads, and one hates all the roads and the other likes all the roads. The result is—well, some things are not hard to calculate. They stand at the cross-roads." (Chesterton, "Orthodoxy," ch. 3).

And that is not much, by way of purpose...

Lee comes to the following position: we freely chose, and not some random force, and specifically not the forces of nature. But machines are quite bound to nature and natural forces, and thus I think they cannot be free.

In many ways I answered this statement in advance. By the Church-Turing thesis and the ability of scientific reductionism to describe our brain completely...
But this is just what I mainly have been arguing about, we can't simply assume this is true. Why is a chimp brain not pretty close in ability to our own? And computers are far from any brain machinery. And how can we know that our actions are solely products of our atoms and their initial positions and velocities (plus random quantum effects)?

If this is all there is, then this is decidedly not what people mean when they say they freely choose. If naturalism is a complete explanation, then people do not in any sense really make a choice, there is even essentially no agent to talk about, if we have no independence from "all there is," you might as well say a summer meadow decided to grow grass, or the sun decided to rise this morning. But none of this is what people mean when they say they made a choice, they mean some independence from other agents, and also freedom from the chain of natural causes.

As indicated by people's reactions when you say "Well, you just decided to do that!" They will protest, and seek to make it plain that they were really involved, as an agent, as a person, choosing deliberately, and with real insight, and also insisting that their choice wasn't merely random.

... as computable functions our minds may – by definition almost – be able to be run on a computer. Where then is the special additional force that is not random and not the forces of nature. In hundreds of years of the growth of scientific knowledge there is not a field of not nature.
That perhaps would be because science concerns itself only with nature! If you never stray from Portugal, you will not land in America. The additional force, I agreed not to talk about! But I believe it is there, maybe I can still approach this somewhat by mentioning the argument from reason: if our thoughts are solely determined by non-rational processes, we have no claim to real insight. There is no real foundation for the validity of our reasoning. But can we be pragmatists? Can we say that "'because a thought is useful, therefore it must be (at least partly) true.' But this is itself an inference. If you trust it, you are once more assuming that very validity which you set out to prove." (C.S. Lewis, "Miracles," ch. 3).

So the validity of reason is an axiom, yet if we start with naturalism, and this axiom, we reach an evident contradiction, we cannot hold that our insights can be essentially reliable, if we are not real agents, and natural forces produce all our thoughts for us, in that case our thoughts about our thinking are simply ordered for us as well! It indeed could (would?) be a selective advantage to think that our thoughts are valid. Never mind whether they are really true or not, we will think we are reasoning correctly, regardless, if evolution dictates that, and thus we can't trust our thoughts about our thoughts.

So we must abandon one of these premises, I prefer to abandon naturalism, instead of saying we have no real assurance of any insight, and thus life is mindless insanity, with a sweet topping of self-deception. Wait, no, that second conclusion is an inference! I couldn't trust that either, if I held that second view...

can a machine have a will?

I tend to consider will as a choice made by a thinking agent and the purpose being the outcome independent of the choice.
Yes, a desired outcome I would hold to be a purpose, yet that implies an agent! And a bird is not usually considered an agent in this sense, nest building is wired into the sparrow, yet a purpose implies a choice, and a sparrow will build a nest regardless. It doesn't choose, nor (I expect) does it contemplate an outcome, it cannot do this, and neither can a machine do this. Or at least I would insist on being shown how it can...

I found myself trying to identify something that ALICE did not know to see if ALICE would learn. ALICE would not. I do view this as a major hurdle to overcome, but not the hurdle Lee considers it to be. Why?

Fundamentally because the range of computational tools that are being developed beggars imagination. We have tools which are capable of doing inference, both simple and temporal inference. Situational calculus provides a tool for describing the world, the lot of context, and a world view as Lee describes it are representable in situational calculus. And better yet, there are already tools for solving these types of problems.
But it seems that situational calculus is quite similar in expressive power to predicate calculus, I don't know myself of an essential advance in the ability to do inference, it seems the problems there are thorny still, otherwise, we would hear of the progress in the discussions of this (for example, here (http://www.illc.uva.nl/Publications/ResearchReports/PP-2001-17.text.pdf) and here (http://en.wikipedia.org/wiki/Situation_calculus)), and see Robby the Robot on the Oprah Winfrey show.

Also, even predicate calculus is undecidable, it has been shown (http://en.wikipedia.org/wiki/First-order_logic) that there is "provably no decision procedure for determining for an arbitrary formula P, whether or not P is valid."

So how can we assure ourselves that logic systems such as these can provide a usable worldview representation, from which we can make our necessary deductions?

And it seems the situation is worse than this, situational calculus seems to have essential limitations in simply describing our worldview, for instance, how can we represent good judgment? How can a description of impressionist paintings (so we can conclude "this is an impressionist painting, this might be, this most certainly is not") be given with functions and predicates? It is tempting to think that such expressions are sufficient, because they cover an infinite space, yet the space seems to lack some necessary descriptive power, it is just not complex enough.

As a second example, consider giving a tip at a restaurant. We leave a tip if the waitress has been helpful, if we have enough money, and maybe also if the food has been good, but the list doesn't end there! If I don't go insane right this minute, if the waitress doesn't go insane, I leave her a tip. If my [putative, I don't have children] two-year-old doesn't run out the door at this moment! If an alien spaceship doesn't land in the parking lot, and start an earthquake. It is impossible to enumerate with axioms all the qualifications for leaving a tip. Maybe the commander of the spaceship tells me "stop gawking and give her a tip, already!"

And we don't go down such a list, checking off possibilities and restrictions one by one. What situational calculus represents as axioms, etc., we arrive at with perception and reasoning, we see if the commander in an alien spaceship were to tell us to give her a tip, we would probably be well advised to do so. But we don't have that on a checklist, we don't tick it off as not preventing us, before putting our dollars on the table.

OTTER is a tool for providing solutions to, and the reasoning behind, how you solve various problems. ... The missing piece at present is the ability to convert the problem from language to the structured form shown.
I, of course, would disagree! If the waitress turns out to be my neighbor's errant daughter, who has been neglectful in serving, but who nevertheless should have some encouragement ... leave a tip. And admire the impressionist painting, on the way out.

Regards,
Lee

[KnightWhoSaysNi]

James T,

Please note that the time limit to post your next statement has expired. The rules, however, permit you a 3 day grace period to post your statement.

Thank you for your consideration,

- NS, FD Moderator

[James T]

Apologies for missing the deadline. I had myself on a Saturday schedule not remembering it was the time between posts. And then I slipped over (badly) the 2000 word limit with casual quoting and too much waffle. For this round I'm running with the more normal thread format.

I would say that reasoning implies real perception, for example, seeing why a point is true, as in the question I always dreaded on exams, "Why or why not?" The teachers asked that, of course, so that the poor students could demonstrate that they really understood <snip>A good point, and we usually expect people who have really understood to be able to apply their knowledge in new situations, not just regurgitate. The fact is that AI systems in use today, expert systems being an example, can already do this, and explain their reasoning process. This was the state of the play 20 years ago when I was at university and doing a paper called Knowledge Engineering, it was unfashionable to call the course AI engineering at the time.

Clearly this illustrates that this is not a problem in principle. We are well past the first steps on all of these processes. Interestingly we started by extending human abilities, now we are backfilling. Almost by definition when the back filling is complete AI will already have surpassed us.

The goal is for perception, not processing, insight, and not just sight, <snip>

So what is human perception? I think this is where the brain touches the supernatural, so I can't really go much further in making a positive case here, since we agree to not talk about God or souls,This is an expression of something I addressed in round 1. Intelligence is not something that artificial intelligence does not have. In terms of the state of technology and the empirical tools, we know that there are roughly 30 visual processing centres in the brain, we know roughly what they do, we can replicate the functionality already. There is nothing special in this we cannot, in principle, do.

<snip> so how can we know that the mechanisms available today can match human reason? And making this conclusion based on future, yet unknown mechanisms and algorithms seems unconvincing to me.As I have stated, we use fMRI and the other empirical tools that we have and will have. These are telling us that there is a rich field of empirical research. The beauty of this is that it is clear there is a lot to learn and that we have the tools to begin studying it. In every field of science that has matured these are the tools that are required for understanding and with the mind, we have them! The amazing thing is how fast we fill the gaps when we have the empirical tools at hand.

Well, can a computer have an out-of-body experience?Sigh.

Descartes had the cool idea to question foundations that were so basic they were assumed. The shame is that he followed the cool idea by such mystical metaphysical waffle. The mind body problem is pretty categorical. Dualism doesn't work, neither being necessary nor sufficient to describe any observable in our universe.

I would refer here to what I mentioned about animal behavior! Why does the chimp's brain not give some substantial fraction of the mental power of a human?Cool pointer lee_merrill, I hadn't thought to look in this area. But I have to say I think this really just supports my position, that AI is coming and we will – by comparison – be the dumb cousins.

From here (http://www.psych.utoronto.ca/~reingold/courses/intelligence/cache/1198yam.html) provides a more scientific view on the field, as do estimates of Koko the gorrilla's IQ at somewhere in the range of 70-95.
But how can you tell if an animal is self-aware? In the late 1960s Gordon G. Gallup, Jr., devised a now classic test using mirrors. Gallup painted a red dot on the faces of anesthetized animals and then observed them when they awoke and noticed themselves in the mirror. An animal that would start poking at the red spot on its face seemingly indicated an awareness that it was seeing itself in the mirror, not another creature. Of all the animals tested in this way, only humans, chimpanzees and orangutans pass.

One main point that I would insist on in reply is that this needs to be proven! Yes, there is no problem in principle,Until such time as it has been done the fact that it has not been done can be used as an argument for why it cannot be.

Similarly you could say that because I have never added 19028946291 and 18162349127 together I cannot. Not very convincing.

John McCarthy has a description of simple deterministic free will in situational calculus. In a sense this level of free will is also available to a chess program.
I'm not sure what you mean here, though, chess programs, as far as I know, have a way of strictly computing the next move, so how is that free will in any way? And what would "deterministic free will" be? Could you provide some more description?Sure, we agreed to define free will as that which a person has. The people in the P forum will argue this point to death without adding AI to it and I really just wanted to demonstrate that there was no problem in principle with AI having the same level of free will as a person. The Church-Turing thesis says we can run anything computable on a Turing Complete computer. There is nothing to demonstrate anything in the brain is not computable, either in a theoretical sense or in modelling a neuron (I understand complete modelling of a single neuron has already been done, there is now a project to model a complete brain). Here, with my McCarthy quote, I was illustrating that some of the conceptual theoretical tools for understanding how to function in a general environment exist. I equivocated with the chess program because in a sense the chess program is functioning in an environment that, while constrained, is not in principle different from a general environment.

... even a successful test, passing the Turing test, will not show that AI exists.I agree! I could be simulated in the area of cooking or car repair by a computer that prints repeatedly, "I don't know," to virtually every question.I used this statement as an introduction to refuting Searle's argument, so this is a bit of a misquote. Perhaps if you addressed my refutation of Searle's argument instead.

This is as if biological phenomena are somehow beyond the materialistic view of the universe.The important question to decide here would seem to be if consciousness is a biological phenomenon, I guess the important part of the Chinese room is whether there can be consciousness in the program that the agent is mechanically carrying out! That would be pretty wild, <snip> I was demonstrating that even Searle was not proposing a dualistic interpretation. Your attribution of pretty wild is assuming that which you wish to prove again. It's not wild if is most reasonably actually is, and the system of a person and Chinese instruction book (assuming it has state information that can change) is most reasonably interpreted as a AI which understands Chinese. Interesting question though, does the person leaving the room kill the Chinese fluent composite?

Have your program check "Blibble is a nonsense word"!A pointer to another nice point, not a demonstration of impossibility again but to more grist for the mill in building our smarter replacements :D. Ramachandran (http://www.bbc.co.uk/radio4/reith2003/lecture1.shtml), search for bouba in this copy of the Reith lecture. Take Bouba and Kiki as names, take two objects, one sharp and one amoeboid. 98% of people favour a particular association between the words and shapes.

The AI bots in these games are coded to feel like human opponents. Many of them are very good.But isn't that like Weizenbaums' Eliza? Really smoke and mirrors?I suppose you make the same accusation of human consciousness. I started a thread in [P] some time ago about the illusion of conscious thought. Once you accept the certainty of AIs arrival, as I have, there are some interesting implications of the Church-Turing thesis, Turing complete computational systems, databases and distributed processing with respect to identity. It's not properly part of the debate but it is a great follow-on idea for a thread.

While you can and may argue that what has not already been done cannot be generally a convincing argument requires more, some reason and justification that demonstrates why it is more credible to believe in the impossibility than the possibility.

Regards James T.

[KnightWhoSaysNi]

The next two concurrent statements will make up the Final Round.

[lee_merrill]

Hi James,

... we usually expect people who have really understood to be able to apply their knowledge in new situations, not just regurgitate. The fact is that AI systems in use today, expert systems being an example, can already do this, and explain their reasoning process.
Yes, but as far as I know, that would simply be theorem provers making a new proof, when presented with new circumstances, and then presenting the proof. But making a new proof is not comprehension, there is no idea in the theorem prover (before setting out on a new proof) whether or not the conclusion is likely or not. They just try and see if they can find a proof of something, or a solution, etc.

But when people say "yes" or "no" on a test, with understanding, they usually have not done a proof! But they do think their view is a real insight, and when asked to defend it, in a new setting, they expect they can do so, because it is a valid insight.

Yet the theorem provers and solution searching programs had no inkling of this from what I had heard of (when I was taking such courses as you mentioned) about 30(!) years ago, and I expect there is still no program that can tell you whether a statement they have proved in one context is probable in another one, before they set out trying to find a proof. I conclude they don't have insight, like people do.

Interestingly we started by extending human abilities, now we are backfilling.
I would attribute the remarkable abilities to specialized algorithms specific to narrowly-focused problems, though, not general ability. Ask your computer chess program to play checkers! Ask the phone call routers to solve a maze...

Lee: The goal is for perception, not processing, insight, and not just sight...

James: Intelligence is not something that artificial intelligence does not have.
Well, that would be your conclusion! It would not, I would say, be mine...

James: In terms of the state of technology and the empirical tools, we know that there are roughly 30 visual processing centres in the brain, we know roughly what they do, we can replicate the functionality already.
Yes, but this is vision, which is unrelated to insight and perception, that is my view (no pun intended) here.

Lee: ... so how can we know that the mechanisms available today can match human reason? And making this conclusion based on future, yet unknown mechanisms and algorithms seems unconvincing to me.

James: As I have stated, we use fMRI and the other empirical tools that we have and will have.
But I need to know what these new tools will be, an MRI does not show us real reasoning, any more than looking at a person's mouth moving tells us of how they are producing reasons in their arguments.

The beauty of this is that it is clear there is a lot to learn and that we have the tools to begin studying it.
Yes, it's fun to explore, I do agree...

James: In every field of science that has matured these are the tools that are required for understanding and with the mind, we have them!
I don't think we do, though! I would like a tool that measures the positions of all the atoms in a given human brain. Only we now know that that is impossible, atoms cannot be measured completely, by definition.

So if we cannot in principle even measure a physical system, we not may be able to fully understand it.

Lee: Well, can a computer have an out-of-body experience?

James: The mind body problem is pretty categorical. Dualism doesn't work, neither being necessary nor sufficient to describe any observable in our universe.
It would seem that there is difficulty in explaining some observables in the non-dualist view, though! So I don't know that we have a way to dismiss either option completely.

James: But how can you tell if an animal is self-aware?
I would say that self-awareness is not a critical aspect of intelligence, though. And also still point to the extraordinary gap between human and great ape abilities. The great apes are not at all great, in that area, and the gap is enormous, yet they have such similar brain machinery, it would seem!

"If you leave off looking at books about beasts and men, if you begin to look at beasts and men then (if you have any humor or imagination, any sense of the frantic or the farcical) you will observe that the startling thing is not how like man is to the brutes, but how unlike he is. It is the monstrous scale of his divergence that requires an explanation. That man and brute are like is, in a sense, a truism; but that being so like they should then be so insanely unlike, that is the shock and the enigma. That an ape has hands is far less interesting to the philosopher than the fact that having hands he does next to nothing with them; does not play knuckle-bones or the violin; does not carve marble or carve mutton." (G.K. Chesterton, "Orthodoxy," ch. 9).

Lee: One main point that I would insist on in reply is that this needs to be proven!
Yes, there is no problem in principle, if people's perception of what they do when they make choices is incorrect, if we are merely mechanical...

James: Until such time as it has been done the fact that it has not been done can be used as an argument for why it cannot be.
Now if you are concluding that it's possible, then of course I have no such strong disagreement with you. But you are claiming more than a possibility, you are concluding that all that happens, including human reason, is completely mechanical. For that I would require good evidence, though I might begin to doubt such evidence if it was itself produced by this machine, controlled by an essentially unreasoning process.

James: The Church-Turing thesis says we can run anything computable on a Turing Complete computer. There is nothing to demonstrate anything in the brain is not computable...
And we also have difficulty demonstrating that everything in the brain is computable! So the Church-Turing thesis doesn't seem to help us here.

I understand complete modelling of a single neuron has already been done, there is now a project to model a complete brain.
A chimp brain? There is a problem here that we may not have addressed even in the best of models...

I equivocated with the chess program because in a sense the chess program is functioning in an environment that, while constrained, is not in principle different from a general environment.
Only I don't think that a chess program has any free will, which is very easily shown to be computable! Not in the sense the people say they have free will, for they don't mean random, they really imply choices independent, in a real way, of natural causes.

Lee: I guess the important part of the Chinese room is whether there can be consciousness in the program that the agent is mechanically carrying out! That would be pretty wild...

James: I was demonstrating that even Searle was not proposing a dualistic interpretation. Your attribution of pretty wild is assuming that which you wish to prove again.
I would say this was argument from astonishment! Can my thoughts themselves have thoughts? Most people would say no on that one, though the situation certainly might be different.

Interesting question though, does the person leaving the room kill the Chinese fluent composite?
I guess that would be a similar question, if my thoughts ever have an independent consciousness, does that then lead to birth and death of personality? And then those thoughts could have thoughts and thus another personality that has thoughts that have ... well, you get the idea.

The world would then have a lot of inhabitants indeed, if that happened. And I think I will leave it at that!

Lee: Have your program check "Blibble is a nonsense word"!

James: A pointer to another nice point, not a demonstration of impossibility again but to more grist for the mill in building our smarter replacements...
But the basic method needs to change in an essential way! Just putting in more code to handle this new special case will not provide us a real advance in intelligence.

Lee: Really smoke and mirrors?

James: I suppose you make the same accusation of human consciousness. I started a thread in [P] some time ago about the illusion of conscious thought.
Actually! I would say that conscious thought is real, and not an illusion, because it is not solely determined by the chain of natural causes. But a thought that was an inevitable product of the bumping of all the atoms in the universe, that thought I would hold to be illusory, if it was thinking it was necessarily a real insight.

James: ... there are some interesting implications of the Church-Turing thesis, Turing complete computational systems, databases and distributed processing with respect to identity.
That would indeed be an interesting topic, though it would be for those you have convinced! I shall not even say I have an identity, though, if there is nothing I can really do, independently of the bumps of all my atoms.

James: .... a convincing argument requires more, some reason and justification that demonstrates why it is more credible to believe in the impossibility than the possibility.
Well, in summary, I would mention the human brain and the apparently quite similar chimp brain, and the enormous gap between them, this makes it seem to me less probable that we can compute our way to human reasoning.

I would also mention the view people have of their free will, that they are not being forced to choose by mechanical processes, and also that their choices are not random. But those are the only two options we know of, in a naturalistic view.

I don't know the limits of what we can explore! And I'm delighted to try and see what computers can do, neural nets are fun, and interesting, so I look forward to what people will come up with in this and similar areas.

And if I find myself refuted in my view on this by a neural net some day, I will be refuted in more sense than one!

Regards,
Lee

[James T]

The only principle that can be applied which might show that Artificial Intelligence is impossible is the fact that it has not already been done. However this is not a compelling argument for at some point in time all the technology we have did not exist.

Will AI surpass us? Almost by definition this is true. We have used computers to extend our capabilities. When we backfill the conscious capabilities these capabilities, already superior to our own, AI will have access to abilities naturally of it's own flesh which we access through keyboard, mouse and computer screen.

In a sense the outcome is not inevitable, in an absolute sense. Perhaps in a relative sense our creation of AI is less certain than that the sun will rise in the morning. However while we exist and progress, people will do what people do and reasearch, experiment and eventually create AI.

Lee The Machine
This story of Lee's efficacy in cranking out code, and the complement from his boss, shows that in society there is a tendency to ascribe positive performance to machine-like behaviour. In a sense society has already accepted that the coming of AI is inevitable and that when it arrives it will set the standard, it will surpass us.

Free Will, Monism and Dualism
A contentious point at the best of times, hence my stipulation that free will be accepted as that which humans have. Accepting neutral monism as the best description of the human mind and using the results of the Church Turing thesis the possibility of simulating the brain in a computer exists. This clearly demonstrates that it is possible for AI to have free will in principle, even if we have to copy a human mind to get there.

Accepting monism is reasonable since empirical observation indicates the brain is both necessary and sufficient for the mind. Dualism fails on the mind body problem, we are unable to conceive of a reasonable explanation for how the non-physical mind can interact with the physical brain. This is further complicated by the fact that for damage to specific parts of the brain, specific parts of the non-physical mind must also fail ... how is this? ... Dualism thus creates problems that do not exist in neutral monism while – at the same time – failing to add any explanatory power on the mind.

The Turing Test
Computers are already capable of fooling some people some of the time with their emulation of human behaviour. They are not yet ready to fool people looking to demonstrate their artificiality. My own test would involve attempting to teach the AI something new. I recall a teacher who described knowing the student well enough to predict the stumbling block in the way of their understanding. Perhaps a professor, with the intimate knowledge of how people fail in learning, might observe a completely new failure to understand. However, I have my doubts, people are incredibly creative in their errors.

Searle's Chinese Room Argument is a disappointing effort to demonstrate impossibility. Disappointing because it begs the question as to whether the being, consisting of an English speaking person and a rule book set, does not itself form an intelligent being. A less fallacious argument would be nice.

Experience qua Experience
Fortunately Frank Jackson provided a more interesting example with the super scientist, who despite knowing all there is to know about seeing colour, learns something new on first seeing colour. However I do find it strange that Mary, the super scientist, fails to also study the experience of new experiences. This then reduces Frank's argument to the absurd position that just because you have not added every two integers together, you learn something entirely new each time you do so with numbers you have not added before.

Science Not Philosophy
With the progress that has already occurred many elements of the mind are moving out of the philosophical realm into the realm of science. Empirical epistemology and other fields if you like. Philosophy has been good at spawning new fields of science and today, we have rich fields of research opening with fMRI and other scanners and the ability to prod with transcranial magnetic stimulation. These fields of reasearch, particularly in cognitive neuroscience, have already shown the way that pieces of our self can be broken off and lost. Mirror agnosia, where left is somehow fundamentally lost by the patient, is a striking example.

The Future of AI
During the course of the debate I have gradually noticed two glaring elements in our typical treatment of AI. One of these is the idea that a supercomputer will be required to run one. The Playstation 3 TFLOPS estimate is close to the required ball park. An AI on a DVD that you power up and down at your whim is not beyond the realms of the reasonable.

The second idea is the prevalence of the concept of an AI being countable. The though that if you have two computers running AI software you have two AIs and when you network them you still have two, my suspicion is that the number of AIs present when the network cable is fitted will depend. Integer addition of beings will not apply.

[KnightWhoSaysNi]

The formal debate is now complete. We would like to thank lee_merrill and James T for their participation. Discussion can be continued in the peanut gallery.

- NS, FD Moderator